Publicação
Synthetic data approach for traffic sign recognition
| Resumo: | Currently, Advanced Driver Assistance Systems (ADAS) have been gradually increasing their presence in everyday life, thanks in part to its ability to recognize several distinct types of objects in the road, namely, traffic signs. These systems employ Convolutional Neural Networks (CNNs), a type of classification algorithms that relies on an enormous amount of data in order to be effective. Current traffic sign datasets suffer from a scarcity of samples due to the necessity of compiling and labeling them manually. Such task is highly resource and time consuming. Thus, researches resort to other mechanisms to deal with this problem, such as increasing the architectural complexity of the neural networks or performing data augmentation. This work addresses the data shortage issue by exploring the feasibility of developing a synthetic dataset. Such set would not require gathering and labelling manually thousands of real word traffic sign images, requiring only easily collectable information and no human intervention. The only data required is a set of templates for each sign given that a particular sign may have more than one template. This is required to cope with outdated pictograms that are still present in streets and roads. We apply several colour and geometric processing methods to the templates aiming to achieve a look similar to real signs, from the CNN point of view. One of such methods is the usage of Perlin noise to both simulate shadows and avoid the clean and homogeneous look that templates have. Two use cases for synthetic data usage are presented: considering the synthetic dataset as a standalone training set, and merging synthetic data with real samples when real data is available. The first option provided results that not only clearly surpass any previous attempt on using synthetic data for traffic sign recognition, but are also encouragingly placing the accuracies obtained close to state-of-the-art results, with much simpler networks. The second approach provided results on three distinct test datasets that consistently beat state-of-the-art results, either in accuracy or in simplicity of the network. |
|---|---|
| Autores principais: | Silva, Diogo Lopes da |
| Assunto: | Synthetic data Traffic sign recognition Convolutional neural networks European traffic signs Dados sintéticos Reconhecimento de sinais de trânsito Redes neuronais convolucionais Sinais de trânsito europeus |
| Ano: | 2019 |
| País: | Portugal |
| Tipo de documento: | dissertação de mestrado |
| Tipo de acesso: | acesso aberto |
| Instituição associada: | Universidade do Minho |
| Idioma: | inglês |
| Origem: | RepositóriUM - Universidade do Minho |
| Resumo: | Currently, Advanced Driver Assistance Systems (ADAS) have been gradually increasing their presence in everyday life, thanks in part to its ability to recognize several distinct types of objects in the road, namely, traffic signs. These systems employ Convolutional Neural Networks (CNNs), a type of classification algorithms that relies on an enormous amount of data in order to be effective. Current traffic sign datasets suffer from a scarcity of samples due to the necessity of compiling and labeling them manually. Such task is highly resource and time consuming. Thus, researches resort to other mechanisms to deal with this problem, such as increasing the architectural complexity of the neural networks or performing data augmentation. This work addresses the data shortage issue by exploring the feasibility of developing a synthetic dataset. Such set would not require gathering and labelling manually thousands of real word traffic sign images, requiring only easily collectable information and no human intervention. The only data required is a set of templates for each sign given that a particular sign may have more than one template. This is required to cope with outdated pictograms that are still present in streets and roads. We apply several colour and geometric processing methods to the templates aiming to achieve a look similar to real signs, from the CNN point of view. One of such methods is the usage of Perlin noise to both simulate shadows and avoid the clean and homogeneous look that templates have. Two use cases for synthetic data usage are presented: considering the synthetic dataset as a standalone training set, and merging synthetic data with real samples when real data is available. The first option provided results that not only clearly surpass any previous attempt on using synthetic data for traffic sign recognition, but are also encouragingly placing the accuracies obtained close to state-of-the-art results, with much simpler networks. The second approach provided results on three distinct test datasets that consistently beat state-of-the-art results, either in accuracy or in simplicity of the network. |
|---|